Module codec 

Residual quantization codec for WARP

This module implements the residual quantization scheme used to compress token embeddings in the WARP algorithm. Each vector is decomposed into:

• A centroid (learned via k-means)
• A residual (difference from centroid), quantized to 2-4 bits per dimension

The codec enables efficient scoring without full decompression by using precomputed centroid scores and bucket weights.

Structs

ResidualCodec

Residual quantization codec for compressing token embeddings.

ResidualCodecBuilder

Builder for creating a ResidualCodec from a WarpIndexConfig.